React with Scratchpad

This codebase demonstrates a way to implement the ReAct (Reason+Action) framework for interacting with Large Language Models (LLMs) to produce better responses to inputs by utilizing a scratchpad to help address the problem of the model hallucinating observations.

See the artificial-brilliance/react repo for a demonstration of the ReAct framework using the Thought/Action/Observation pattern.

The code is inspired from a mixture approaches from the following along with custom tweaks and additions:

• The original ReAct paper
• Simon Willison's blog
• LangChain

Usage

This repository was written using Python 3.11 and uses the pdm tool to handle dependencies.

To get started

1. Clone the repo:

     git clone git@github.com:artificial-brilliance/inverted_react.git
   

2. At the root of the repo, install necessary dependencies:

     pdm install
   

3. At the root of the repo, run the code with:

     pdm run start '<some question to answer>'
   

4. (Optionally) run tests using:

     pdm run test
   

Examples

The following example shows what happens when asking the LLM a question that it cannot know because (at the time the code was run) the iphone 15 was not released yet and descriptions of its release date were (most-likely) not in any training data.

$ pdm run start 'when was the iphone 15 released'
> System Prompt:
You are an assistant helping a human.

Your task is to answer the following:
when was the iphone 15 released

You use a Thought/Action/Action Input/Observation pattern:
'''
Thought: The thought you are having based on the observation
Action: The action you want to take.  One of [search]
Action Input: The input to the action
Observation: An observation
'''

(...this Thought/Action/Action Input/Observation pattern can repeat N times)

== TOOLS ==

You have access to the following tools:


Tool 1:
Action: search
Action Input: The query used to search wikipedia
Description: Used to search wikipedia



== OPTIONS ==

Select ONLY ONE option below:

Option 1: To use a tool, you MUST use the following format:
'''
Thought: Do I need to use a tool? Yes because of the observation [the reason why you need to use a tool...]
Action: MUST be one of [search]
Action Input: The input to the tool
Observation: The observation from the action
'''

Option 2: To output the final answer, you MUST use the following format:
'''
Final Answer: The final answer for the task
'''



> Scratchpad:

> Running: search "iPhone 15 release date"
* Cache hit for query ""iPhone 15 release date""
> Scratchpad:

Thought: Do I need to use a tool? Yes, because I don't know the release date of the iPhone 15.
Action: search
Action Input: "iPhone 15 release date"
Observation: Friday, September 15 (updated) Eastern and 1 p.m. U.K. Despite persistent rumors that the iPhone 15 Pro Max would be delayed. it's due for September 22 as well, though I think it's possible that it may be in short supply, so prompt pre-ordering is suggested to avoid delays.Sep 12, 2023

Final Answer: The iPhone 15 was released on September 22, 2023.